Re-write training docs

[mcabbott]

This re-writes training.md to move from implicit to explicit, #1986.

And creates a new page train_api.md to hold the docstrings. For now they live together but the goal is that they can live in different sections, "narrative vs. lookup" -- see #2115

Also updates basic.md to introduce explicit gradients first.

In all of these, it puts implicit things in !!! note boxes nearby. I start to think that's better than having a separate page for implicit-to-explicit changes.

It expects #2082 in order for update! and Adam to work with explicit gradients, and for the new train!. Rebased. But struggling to make the right docstrings show up. One update! signature isn't co-operating.

After this I think regularisation.md needs some re-working. It could be a home for some of what's in optimisers.md (except the listing of rules), basically all the more advanced topics after this? I.e. this page is the basic idea of what a training loop looks like; the next one covers regularisation, freezing, scheduling, etc. Or maybe it should just be one page? This gets all the blue boxes telling you what to change in one place.

[github-actions]

Once the build has completed, you can preview any updated documentation at this URL: https://fluxml.ai/Flux.jl/previews/PR2114/ in ~20 minutes

Edit: especially https://fluxml.ai/Flux.jl/previews/PR2114/training/training/

Compare to http://fluxml.ai/Flux.jl/stable/training/training/ + http://fluxml.ai/Flux.jl/stable/models/regularisation/

[mcabbott]

I think this is about ready.

It could use another set of eyeballs to see whether anything from the old docs has been lost, or made more confusing.